No problems or questions, it just seems rare that anyone shares what they're working on...
I've been using DL to experiment with new kinds of architectural drawings. Currently using a deeper DCGAN to generate architectural plans. Here are some results after ~20,000 iterations.
Obviously a little sketchy but the original paper starts getting very good results around 100,000 iterations.
The network is almost identical to the DCGAN example, but now with 9 convtranspose layers (~100mil parameters for G and D) from z up to 3x512x512
Thanks for making pytorch fun & easy